|
| 基于机器学习的非心肺转流冠状动脉旁路移植术相关急性肾损伤的预测模型 |
| Establishment of a predictive model for acute kidney injury related to off-pump coronary artery bypass grafting based on machine learning |
| |
| DOI：10.12089/jca.2023.05.001 |
| 中文关键词: 非心肺转流冠状动脉旁路移植术 急性肾损伤 机器学习 可解释性模型 |
| 英文关键词: Off-pump coronary artery bypass grafting Acute kidney injury Machine learning Shapley additive explanations method |
| 基金项目:辽宁省重点研发计划项目（2019JH8/1030083） |
|
| 摘要点击次数: 997 |
| 全文下载次数: 394 |
| 中文摘要: |
目的 建立基于机器学习的非心肺转流冠状动脉旁路移植术相关的急性肾损伤（OPCABG-AKI）可解释性机器学习预测模型。
方法 回顾性收集2018—2021年行OPCABG的1 110例患者的临床资料。建立并比较8种机器学习模型，采用Python的SHAP模型解释包对预测性能最佳的黑箱模型进行解释性分析。将特征参数SHAP绝对值的平均值定义为该参数的重要性并进行排序；以SHAP值为依据确定各特征参数与OPCABG-AKI的关系；对主要风险因素进行单个特征量化分析；对模型中具有代表性的真阳性及真阴性样本进行独立的解释性分析。
结果 共有405例(36.5%)患者发生AKI。在8种机器学习模型中，随机森林（RF）预测模型性能最优，针对阳性样本的受试者工作特征曲线(ROC)下面积(AUC)为0.90(95%CI 0.86~0.94)。SHAP模型解释性分析结果显示术中尿量对RF模型的贡献最大，其次为诱导期循环变异系数、术中右美托咪定用量、术中舒芬太尼用量、术中低血压时间、术前血清肌酐基线、APACHE Ⅱ分数和年龄等。
结论 以随机森林集成学习算法构建模型可较好地预测OPCABG-AKI，模型中术中尿量等指标与OPCABG-AKI关系密切。 |
| 英文摘要: |
Objective To establish an explanatory prediction model of machine learning for the acute kidney injury related to off-pump coronary artery bypass grafting (OPCABG-AKI) based on the machine learning.
Methods The clinical data of 1 110 patients who underwent OPCABG from 2018 to 2021 was collected retrospectively. Eight models of machine learning were established and compared, and the SHAP model explanation package of Python was used to conduct the explanatory analysis of the black box model with the optimal prediction performance. The average value of the absolute SHAP values of characteristic parameters was defined as the importance of the parameter and sorted them; the positive/negative relationship between each characteristic parameter and OPCABG-AKI based on the SHAP value was determined; for the major risk factors, the quantitative analysis of single feature were conducted; and the representative true positive and true negative samples in the models were also conducted.
Results A total of 405 patients (36.5%) had AKI. The performance of the random forest (RF) prediction model was the best among 8 models of machine learning. The AUC for positive samples was 0.90 (95% CI 0.86-0.94). The explanatory analysis of the SHAP models showed that the urine volume during operation has the most contribution to the RF model, and the others were the coefficients of cyclic variation during induction, such as the dosages of dexmedetomidine and sufentanil during operation, the duration of hypotension in operation, the baseline of serum creatinine before operation, APACHE Ⅱ score, and age.
Conclusion Using the ensemble learning algorithm of RF to construct models can predict OPCABG-AKI well, and the urine volume during operation, and other indicators in the models are closely related to OPCABG-AKI. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
